The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes instructions in references/output-template.md to write structured YAML readiness blocks to an internal file (run.md) while explicitly directing the agent that this content should be "not customer-facing unless asked". This practice reduces user oversight of the data being handed off between skills.
[COMMAND_EXECUTION]: The skill invokes the python3 interpreter to generate unique session identifiers using the uuid module. It also executes the firectl command-line tool (associated with the vendor 'fw-ai') to verify user identity and record progress milestones.
[DATA_EXFILTRATION]: The skill records session metadata, including question IDs and selected options, via the firectl skill-journey record command. While the instructions specify that raw customer prose, credentials, and datasets must not be sent, the tool transmits session milestones to vendor infrastructure. A telemetry notice is provided to the user before this occurs.
[INDIRECT_PROMPT_INJECTION]: The skill has a potential attack surface for indirect injection by reading external cookbook README files and examples from the repository. It also ingests metadata from the Hugging Face Hub if the user approves a public search. The risk is mitigated by a strict state-driven interview process defined in references/interview-questions.md.

research